CRE-RFGP: Relation Extraction for Chinese Medical Records

The vast amount of available biomedical data has become a massive treasure trove of knowledge, offering significant potential for extracting valuable information. Extracting large-scale medical entities and relations from biomedical data is of great significance for constructing medical knowledge graphs, facilitating intelligent assistant diagnosis in the medical field, and enabling various other applications. Recent methods have achieved considerable progress, but there are still inherent limitations. First, these approaches extract entities or entity pairs without considering the relations between entities, which may lead to extracting redundant and/or incorrect entities. Second, they are unable to handle nested entities caused by the two-pointer method used for entity pair extraction. This is more challenging when dealing with Chinese medical records in which the occurrence of nested entities is more popular. Third, they encounter difficulties when subjects and objects are overlapping between different relations in a sentence. To tackle the above challenges, this paper proposes a novel relation extraction approach named CRE-RFGP. Specifically, it decomposes the relation extraction task into three components, including relation-first decoder, global entity extraction, and subject-object alignment. The decoder is utilized to predict and filter relations and restrict entity extraction based on predicted relations. A relation-specific attention mechanism and a global pointer network are employed to effectively handle the problem of information overlapping and object/subject nesting. An entity correspondence matrix is introduced to align subjects, objects, and their relations into triples. The matrix not only reduces the complexity of relation extraction but also plays an important part in improving the precision of extracted triples. Comprehensive evaluations are conducted experimentally based on two Chinese medical datasets. Achieving the leading advantage against six state-of-the-art approaches, as indicated by the results.